当前位置：首页 > news >正文

LMS/NLMS最小均值算法：双麦克风降噪

news 2025/8/7 16:56:46

NLMS学习率不知道怎么调，效果没LMS好。

原理参考：
https://stc.fs.cvut.cz/pdf17/6563.pdf

测试内容：
noise-16k16bits.pcm是噪音音频文件
mixervoice-16k16bits.pcm 是由演讲录音与noise-16k16bits.pcm按0.5:0.5 比例混合的音频，语音+噪音。

// test-lms-denoiser.cpp : 此文件包含 "main" 函数。程序执行将在此处开始并结束。
//#include <iostream>#include <stdio.h>
#include <stdlib.h>
#include <math.h>#define FILTER_LEN 32    // 滤波器长度
#define MU 0.01f         // 收敛因子（步长）// LMS滤波器结构体
typedef struct {float coeffs[FILTER_LEN];  // 滤波器系数float buffer[FILTER_LEN];   // 输入信号缓冲区int index;                 // 缓冲区当前索引
} LMSFilter;// 初始化滤波器
void LMS_Init(LMSFilter* filter) {for (int i = 0; i < FILTER_LEN; i++) {filter->coeffs[i] = 0.0f;  // 初始系数设为0filter->buffer[i] = 0.0f;   // 清空缓冲区}filter->index = 0;
}// LMS自适应滤波（单次迭代）
float LMS_Update(LMSFilter* filter, float input, float desired) {// 将新样本存入缓冲区filter->buffer[filter->index] = input;// 计算滤波器输出（卷积）float output = 0.0f;for (int i = 0; i < FILTER_LEN; i++) {int j = (filter->index - i + FILTER_LEN) % FILTER_LEN;output += filter->coeffs[i] * filter->buffer[j];}// 计算误差float error = desired - output;// 更新滤波器系数for (int i = 0; i < FILTER_LEN; i++) {int j = (filter->index - i + FILTER_LEN) % FILTER_LEN;filter->coeffs[i] += MU * error * filter->buffer[j];}// 更新缓冲区索引filter->index = (filter->index + 1) % FILTER_LEN;return output;
}// 生成测试信号（正弦波+噪声）
float generate_input(int n) {float signal = 0.3f * sinf(2 * 3.1415926f * n / 50.0f);float noise = 0.1f * ((float)rand() / RAND_MAX - 0.5f);return signal + noise;
}// 模拟未知系统（待辨识的FIR系统）
float unknown_system(float* buf) {float known_coeffs[FILTER_LEN] = { 1,/* 假设的未知系数 */ };for (int i = 0; i < FILTER_LEN; i++) {known_coeffs[i] = 1.0f;}float output = 0.0f;for (int i = 0; i < FILTER_LEN; i++) {output += known_coeffs[i] * buf[i];}return output;
}int main_() {LMSFilter filter;LMS_Init(&filter);float input_buf[FILTER_LEN] = { 0 };int buf_idx = 0;// 迭代1000次训练for (int n = 0; n < 1000; n++) {// 生成输入信号float x = generate_input(n);input_buf[buf_idx] = x;buf_idx = (buf_idx + 1) % FILTER_LEN;// 获取未知系统的期望输出float d = unknown_system(input_buf);// LMS更新float y = LMS_Update(&filter, x, d);// 打印误差（观察收敛）if (n % 100 == 0) {printf("Step %d: Error = %.4f\n", n, d - y);}}// 打印学习到的系数printf("\nLearned Coefficients:\n");for (int i = 0; i < FILTER_LEN; i++) {printf("%.4f ", filter.coeffs[i]);}return 0;
}/// LMS自适应滤波基本参数
#define LMS_FILTER_SIZE 32      //定义FIR滤波器的阶数。为保证高速运算须为8的倍数。
#define LMS_NIU         0.01    //LMS收敛因子
// NLMS归一化自适应滤波基本参数
#define NLMS_MU         0.01f  // NLMS 学习率
#define NLMS_EPS        0.001   // NLMS 防止除数为零的常量// LMS滤波采样数据类型
typedef float lms_node_t, nlms_node_t;/// <summary>
/// LMS对象结构
/// </summary>
typedef struct lms_{lms_node_t coeffs[LMS_FILTER_SIZE];lms_node_t buffer[LMS_FILTER_SIZE+1];lms_node_t error;
} lms_t, nlms_t;/// <summary>
/// 初始化LMS滤波对象
/// </summary>
/// <param name="lms"></param>
/// <returns></returns>
lms_t* lms_init(lms_t* lms)
{if (lms == NULL) {return NULL;}memset(lms->coeffs, 0, sizeof(lms_t::coeffs));memset(lms->buffer, 0, sizeof(lms_t::buffer));lms->error = 0;return lms;
}/// <summary>
/// 对一个信号采样进行LMS滤波自适应，估计滤波后的值。降噪时：假设mic1(desired)=语音+噪声，mic2(input)=参考噪声 mic2_noise, mic1_mixed
/// </summary>
/// <param name="lms">LMS滤波对象</param>
/// <param name="input">降噪时对应：input - 参考噪声 mic2_noise</param>
/// <param name="desired">降噪时对应：desired - mic1为语音+噪声</param>
/// <returns>降噪时对应：噪声估计；lms->error 为干净声音样本</returns>
lms_node_t lms_update(lms_t* lms, lms_node_t input, lms_node_t desired)
{lms_node_t* coeffs = NULL;lms_node_t* buffer = NULL;lms_node_t estimated_noise = 0;lms_node_t coeff = 0;int idx = 0;if (lms == NULL) {return NULL;}coeffs = lms->coeffs;buffer = lms->buffer;buffer[LMS_FILTER_SIZE - 1] = input;// 卷积for (idx = 0; idx < LMS_FILTER_SIZE; ++idx) {estimated_noise += coeffs[idx] * buffer[idx];}// 求误差lms->error = desired - estimated_noise;// 更新参数coeff = LMS_NIU * lms->error;for (idx = 0; idx < LMS_FILTER_SIZE; ++idx) {// 更新滤波参数coeffs[idx] += coeff * buffer[idx];// 缓存前移，挪出下一个采样的位置buffer[idx] = buffer[idx+1];}return estimated_noise;
}/// 归一化LMS滤波初始化
#define nlms_init lms_init/// <summary>
/// 对一个信号采样进行NLMS滤波自适应，估计滤波后的值。降噪时：假设mic1(desired)=语音+噪声，mic2(input)=参考噪声 mic2_noise, mic1_mixed
/// </summary>
/// <param name="nlms">NLMS滤波对象</param>
/// <param name="input">降噪时对应：input - 参考噪声 mic2_noise</param>
/// <param name="desired">降噪时对应：desired - mic1为语音+噪声</param>
/// <returns>降噪时对应：噪声估计，nlms->error 为干净声音采样</returns>
nlms_node_t nlms_update(nlms_t* nlms, nlms_node_t input, nlms_node_t desired)
{nlms_node_t* coeffs = NULL;nlms_node_t* buffer = NULL;nlms_node_t estimated_noise = 0;nlms_node_t power = NLMS_EPS;nlms_node_t coeff = 0;int idx = 0;if (nlms == NULL) {return NULL;}coeffs = nlms->coeffs;buffer = nlms->buffer;buffer[LMS_FILTER_SIZE - 1] = input;// 卷积for (idx = 0; idx < LMS_FILTER_SIZE; ++idx) {estimated_noise += coeffs[idx] * buffer[idx];// 求平方和，计算输入信号功率power += buffer[idx] * buffer[idx];}// 求误差nlms->error = desired - estimated_noise;// 更新参数if (power == NLMS_EPS)power = 0.0f;coeff = NLMS_MU * nlms->error / (NLMS_EPS + power);for (idx = 0; idx < LMS_FILTER_SIZE; ++idx) {// 更新滤波参数coeffs[idx] += coeff * buffer[idx];// 缓存前移，挪出放置下一个采样的位置buffer[idx] = buffer[idx + 1];}return estimated_noise;
}#include <fstream>
#include <vector>void mixerNoiseToCleanVoice();int main()
{std::cout << "Hello World!\n";// 生成混合噪音音频
#if 0{mixerNoiseToCleanVoice();return 0;}
#endif// c++ 版本 LMS 降噪
#if 0{std::ifstream mixervoice("mixervoice-16k16bits.pcm", std::ios::binary | std::ios::in);std::ifstream noise("noise-16k16bits.pcm", std::ios::binary | std::ios::in);std::vector<short> noiseBuffer(1024);std::vector<short> mixerBuffer(1024);std::ofstream clean("output-clean-16k16bits.pcm", std::ios::binary | std::ios::out);std::vector<short> cleanBuffer(1024);LMSFilter filter;LMS_Init(&filter);float input_buf[FILTER_LEN] = { 0 };int buf_idx = 0;while (!mixervoice.eof()) {noise.read((char*)noiseBuffer.data(), noiseBuffer.size() * sizeof(short));mixervoice.read((char*)mixerBuffer.data(), mixerBuffer.size() * sizeof(short));if (noise.eof() || mixervoice.eof()) {break;}for (int i = 0; i < 1024; i++) {float sn = noiseBuffer[i] / 32768.0f;float sm = mixerBuffer[i] / 32768.0f;float x = sn;float d = sm;input_buf[buf_idx] = x;buf_idx = (buf_idx + 1) % FILTER_LEN;// LMS更新float y = LMS_Update(&filter, x, d);cleanBuffer[i] = y * 32768.0f;}clean.write((const char*)cleanBuffer.data(), cleanBuffer.size() * sizeof(short));}return 0;}
#endif// LMS 降噪测试
#if 1{std::ifstream mixervoice("mixervoice-16k16bits.pcm", std::ios::binary | std::ios::in);std::ifstream noise("noise-16k16bits.pcm", std::ios::binary | std::ios::in);std::vector<short> noiseBuffer(1024);std::vector<short> mixerBuffer(1024);std::ofstream clean("output-1-clean-16k16bits.pcm", std::ios::binary | std::ios::out);std::vector<short> cleanBuffer(1024);lms_t lms;lms_init(&lms);while (!mixervoice.eof()) {noise.read((char*)noiseBuffer.data(), noiseBuffer.size() * sizeof(short));mixervoice.read((char*)mixerBuffer.data(), mixerBuffer.size() * sizeof(short));if (noise.eof() || mixervoice.eof()) {break;}for (int i = 0; i < 1024; i++) {float sn = noiseBuffer[i] / 32768.0f;float sm = mixerBuffer[i] / 32768.0f;float estimated_noise = lms_update(&lms, sn, sm);cleanBuffer[i] = lms.error * 32768.0f;}clean.write((const char*)cleanBuffer.data(), cleanBuffer.size() * sizeof(short));}}
#endif// NLMS 降噪测试
#if 1{std::ifstream mixervoice("mixervoice-16k16bits.pcm", std::ios::binary | std::ios::in);std::ifstream noise("noise-16k16bits.pcm", std::ios::binary | std::ios::in);std::vector<short> noiseBuffer(1024);std::vector<short> mixerBuffer(1024);std::ofstream clean("output-2-clean-16k16bits.pcm", std::ios::binary | std::ios::out);std::vector<short> cleanBuffer(1024);nlms_t nlms;nlms_init(&nlms);while (!mixervoice.eof()) {noise.read((char*)noiseBuffer.data(), noiseBuffer.size() * sizeof(short));mixervoice.read((char*)mixerBuffer.data(), mixerBuffer.size() * sizeof(short));if (noise.eof() || mixervoice.eof()) {break;}for (int i = 0; i < 1024; i++) {float sn = noiseBuffer[i] / 32768.0f;float sm = mixerBuffer[i] / 32768.0f;float estimated_noise = nlms_update(&nlms, sn, sm);cleanBuffer[i] = (sm - estimated_noise) * 32768.0f;}clean.write((const char*)cleanBuffer.data(), cleanBuffer.size() * sizeof(short));}}
#endifreturn 0;
}/// <summary>
/// 混合噪音与干净音频
/// </summary>
void mixerNoiseToCleanVoice()
{std::ifstream noise("noise-16k16bits.pcm", std::ios::binary | std::ios::in);std::ifstream clean("clean-16k16bits.pcm", std::ios::binary | std::ios::in);std::ofstream mixervoice("varvoice-16k16bits.pcm", std::ios::binary | std::ios::out);std::vector<short> noiseBuffer(1024);std::vector<short> cleanBuffer(1024);std::vector<short> mixerBuffer(1024);while (!clean.eof()) {noise.read((char*)noiseBuffer.data(), noiseBuffer.size() * sizeof(short));clean.read((char*)cleanBuffer.data(), cleanBuffer.size() * sizeof(short));if (clean.eof()) {break;}for (int i = 0; i < 1024; i++) {mixerBuffer[i] = noiseBuffer[i] * 0.5 + cleanBuffer[i] * 0.5;}mixervoice.write((const char*)mixerBuffer.data(), mixerBuffer.size() * sizeof(short));}
}